How predictions are calculated

Each of the 48 finalists is seeded from its World Football Elo rating — the same family of ratings used by eloratings.net, built from decades of international results. A stronger team carries a higher number; the difference between two ratings is what drives every probability on the site.

The three 2026 co-hosts (USA, Mexico, Canada) get a home-field bump of +87.5Elo points whenever they play — close to eloratings.net’s standard ~100, tuned down for the strength-compressed World Cup field, and worth roughly 16% of extra win probability between otherwise even sides. The bump travels with the host flag; it is never baked into the stored rating.

After each completed match we fold the result back into the two ratings with the classic World Football Elo update:

K = 45 sets how much one game moves a rating; the goal-difference multiplier means a 4–0 shifts ratings more than a 1–0. Because the move is proportional to W − We, an upset (beating a stronger side) is rewarded far more than an expected win. Updates are applied in true kick-off order, so no match is ever scored using information from its own or a later result.

The chance team A beats team B is the standard Elo logistic on the rating difference:

The scalecontrols how sharply a rating gap turns into confidence — a smaller scale makes the favourite’s edge steeper. Mondial uses two scales on purpose:

• 300 — the rating-system scale, used to update ratings after results (and to calibrate against decades of all internationals).
• 500 — the flatter World Cup display scale, used for every probability you actually see (bracket, match cards, title odds).

Why flatter for the World Cup? Because only qualified sides are here, the field is strength-compressed and single matches are high-variance (fatigue, neutral venues, cagey knockouts), so favourites win less often than a curve fitted to friendly-heavy data implies. This was found empirically: tuned on the 2022 World Cup, the flatter scale also improved out-of-sample predictions of the already-played 2026 games. The effect on a single match:

A clear favourite that the rating curve would call ~82% is shown as the more honest ~72%.

For matches that can be drawn we use the Davidson model, which adds a draw term to the Elo logistic:

The draw weight ν = 0.8 is calibrated so two evenly-matched teams draw about 29% of the time. Conditional on a decisive result, the model collapses exactly back to the two-way win probability above, so the group and knockout views always agree. Knockout ties, which can’t end level, simply drop the draw term (extra time / penalties decide).

Each side’s expected goals come from a rating-aware Poisson model: a base scoring rate scaled up or down by the rating gap.

Two independent Poissons under-count low draws, so we apply a Dixon–Coles low-score correction (ρ = -0.03) that nudges scoreline mass toward 0–0 and 1–1 and away from 1–0 / 0–1. The full grid is then conditioned on the calibrated win/draw/away split from step 4, which fixes the overall draw rate — so Dixon–Coles reshapes the scores within each outcome region rather than changing how often a draw happens, and the most-likely scores always line up with the win probability shown elsewhere. The same grid powers the over/under and both-teams-to-score numbers; in knockout (decisive) mode the draw scorelines are removed and the rest renormalised.

Group standings are ranked by the usual ladder — points → goal difference → goals scored → rating (rating stands in for FIFA’s head-to-head steps as a deterministic deep tie-break). From the 12 four-team groups, the knockout field of 32 is:

The eight best third-placed teams are then compared across all 12 groups by points → goal difference → rating (this cross-group ranking skips the goals-scored step the in-group ladder uses). Until a group has played all its games, the model shows each team as a live candidate to finish 1st / 2nd / 3rd rather than a certainty.

The 2026 knockout bracket is fixed by group position: the Round of 32 pairs specific winner / runner-up / best-third slots, so two teams from the same group can’t meet early and each winner’s and runner-up’s path is the official one. The eight best-third slots are filled by a deterministic constrained matching that keeps every structural rule (no same-group R32, correct winner/runner-up slots) but, when several assignments are valid, may place a third-placed team in a different one of its eligible slots than FIFA’s Annex C lookup. Rating only feeds each tie’s win probability, never the seeding.

Each matchup then resolves in priority order:

1. A real result, once played, replaces the prediction and locks in green.
2. Otherwise your pick, if you’ve overridden the tie (saved on your device).
3. Otherwise the model advances the higher win probability, propagating winners round by round to a predicted champion.

A single bracket only shows the most-likely path. To get each team’s chance of reaching every round we run a Monte Carlo simulation — thousands of full tournaments:

1. Sample every unplayed group game (Davidson outcome, then a Poisson margin consistent with it) and rebuild the tables.
2. Work out who qualifies, and slot the official bracket.
3. Play the knockouts as weighted coin-flips on each tie’s win probability, through to a champion.
4. Tally how often each team escapes its group, reaches the final and lifts the trophy.

Already-finished results are held fixed in every run, and the random draw is seeded from the current results state — so the odds are deterministic (the same state always gives the same numbers) and only move when real results do.

The model report card scores every call from only what was known before each match, using log-loss and Brier score (both reward being confident and right, punish being confident and wrong) against a no-skill baseline of ln 3 ≈ 1.099. A reliability table checks that, when the model says 70%, those teams really win about 70% of the time.

The constants above weren’t guessed. The goal model was fit on ~8,000 pre-2022 internationals and checked out-of-sample on the held-out 2022 World Cup; the World Cup flattening was tuned on that 2022 tournament and then validated out-of-sample on the already-played 2026 games. Each was graded on results it was never fitted to.

We also checked whether a fancier model would do better. Every alternative below was trained on the same leakage-free history and scored out-of-sample — on a broad 2023+ holdout and, decisively, on the World Cup matches the app actually predicts. The bar to beat is the shipped Elo + Davidson model, at a log-loss of ≈ 0.882 (lower is better).

Broad-holdout log-loss, versus the shipped model’s ≈ 0.882. Even the few that edge it there don’t transfer to the World Cup split — the field is compressed and the sample is small, so the extra complexity just fits noise.

We also tried richer inputs, not just richer models:

• Match-type features (friendly vs qualifier vs finals) — no measurable gain, and they hurt the World Cup splits; the Elo rating already absorbs match context.
• Recent form, rest days, low-data shrinkage — real but tiny on the broad corpus, and they don’t transfer: the World Cup field is uniformly strong, so shrinkage only removes real rating signal.
• Alternative pre-tournament ratings (corpus-rolled Elo, FIFA-ranking blends, confederation shrinkage) — the registry’s World Football Elo graded best; corpus-rolled Elo was significantly worse, and no blend recovered the gap.
• FIFA world ranking as the rating — a known weaker predictor than Elo.
• xG, squad market value, club-Elo ensembles — no free national-team xG feed, and the rest add friction without a reliable edge.

Across all of it, nothing reliably beat the calibrated Elo + Davidson model on World Cup matches, so we kept the simple, transparent one. The one outside signal that genuinely sharpens predictions is betting-market odds, which the app blends in for upcoming matches when they’re available.

Fixtures, groups and final results come from the public-domain openfootball dataset; live in-play scores, the match minute, the goal timeline and the bookmaker match odds behind the market blend (a DraftKings 1X2 quote, de-vigged) are overlaid from ESPN’s free API — with an optional multi-book consensus from The Odds API outranking that single-book quote when an API key is configured. Squads are enriched from TheSportsDB when the roster looks complete; starting line-ups come from ESPN once a match kicks off, with TheSportsDB (and, failing that, a generated XI) filling any gaps. If a source is unavailable the app falls back to illustrative data and labels it as such, so the model always has something to reason over.

Predictions are driven by team strength and results alone. The model does notknow about specific injuries or suspensions, the announced starting XI, weather, travel and rest, or a team’s motivation once it has already qualified. It assumes ratings capture current strength and that matches are independent. Treat every number as a calibrated estimate of uncertainty — not a forecast of certainty. That is exactly why upsets are not bugs: a 30% underdog is supposed to win roughly three times in ten.